Is this a reasonable graph for the losses?
sarmientoj24 opened this issue · 0 comments
sarmientoj24 commented
I am using the same thing for 8 GPU
base_lr: 0.6
batch_size: 32
checkpoint_freq: 25
crops_for_assign: [0, 1]
data_path: /home/XXX/swav/swav/data
dist_url: env://
dump_checkpoints: dump/test2_ep400_b32_sz_320_128/checkpoints
dump_path: dump/test2_ep400_b32_sz_320_128
epoch_queue_starts: 15
epochs: 400
epsilon: 0.05
feat_dim: 128
final_lr: 0.0006
freeze_prototypes_niters: 5005
gpu_to_work_on: 0
hidden_mlp: 2048
is_slurm_job: False
local_rank: 0
max_scale_crops: [1.0, 0.2]
min_scale_crops: [0.2, 0.1]
nmb_crops: [2, 6]
nmb_prototypes: 3000
queue_length: 3840
rank: 0
seed: 31
sinkhorn_iterations: 3
size_crops: [280, 120]
start_warmup: 0
sync_bn: pytorch
syncbn_process_group_size: 8
temperature: 0.1
use_fp16: True
warmup_epochs: 0
wd: 1e-06
workers: 10
world_size: 8